The AI-infrastructure supply chain: own the connection, not the chip

A data-centre campus that joins the Northern Virginia interconnection queue in 2026 cannot realistically draw utility power before the early 2030s, however fast the building goes up or however much capital the operator commits. The scarce asset in AI infrastructure is the connection and the long-lead equipment behind it, not the chip, and that constraint has already reached South East Asia.

The market reads the AI buildout as a chips-and-capex story, and on direction it is broadly right. The International Energy Agency frames the largest technology firms' capital spending as above USD 400bn in 2025, with a further rise of roughly 75% expected in 2026; on analyst aggregation, the four largest hyperscalers alone are guiding toward roughly USD 630-725bn of capital spending next year, against about USD 380-410bn in 2025.¹² Those are two cuts of an overlapping universe, all large tech in the first and the top four in the second, not two independent confirmations. Power gets acknowledged, but mostly as a macro headline about rising electricity demand. Our read sits one layer below that. The binding constraint is neither chips nor power in aggregate. It is the specific long-lead kit and the in-service grid connection that turn a megawatt into a delivered, dispatchable megawatt at the rack: large power transformers, generator step-up units, high-voltage switchgear, gas turbines, and the interconnection slot itself. Those items set the clock, and value is accruing to whoever already holds them.

What is actually scarce in the AI buildout

The headline asset is abundant or fundable; the gating asset is not. GPUs ship in months and are fungible. Buildings take months to a year and can be financed. The slow, hard-to-replicate nodes sit between the substation and the rack: large power transformers and generator step-up units, reported at roughly two to four-plus years in a Q2 2025 industry survey, with high-capacity units at the long end; gas-turbine slots reserved years ahead; and the interconnection slot itself.³ Value settles at the slowest, least-replicable node, not the fastest. The clearest hard evidence is the queue. Generators that reached commercial operation in the United States in 2024 spent a median of about 55 months, roughly four and a half years, in the generation-interconnection queue, up from about 22 months in 2008.⁴ That is the queue for power plants seeking to connect, not for data centres seeking load, but the two clear the same congested grid through the same study process, so the generator wait proxies the load-side delivery constraint. Capital can buy a campus and a year of chips long before it can buy a connected megawatt.

Most of what joins the queue is never built

Being in the queue is not holding a connection, and the completion record proves it. Of the generation and storage capacity that entered US interconnection queues between 2000 and 2019, only about 19% of projects, and 13% of capacity, had reached commercial operation by the end of 2024; roughly 77% of that capacity withdrew, the balance still active or under construction.⁴ The active queue is large, around 2,290 GW seeking connection at the end of 2024, but it is shrinking, down about 12% on the prior year as some 700 GW withdrew against roughly 500 GW of new requests.⁵ That is fewer connections clearing, not faster ones. The queue is thinning on withdrawals rather than throughput, so an in-service connection is rare and getting rarer against the buildout it has to serve.

Reading the chain as an operator, not a buyer of headlines

Orofante is a family-office investment group that sits inside the energy and infrastructure supply chain rather than above it. We assess turbine and transformer lead times, interconnection economics, and offtake structures as routine work, and we read the AI buildout the way an operator reads any power project: follow the slot, the kit, and the connection, not the headline capex. The analytical unit is whoever already holds the scarce input, not whoever spends the most. That edge is an inventory-and-lead-time position, the kind capital alone cannot shortcut. The same logic anchors our read on wind repowering, where the scarce inputs are land, consent, and an existing grid connection rather than the turbine itself. One discipline matters here. A connection that takes years to win and then earns for decades is structural scarcity; an auction clearing-price spike is cyclical. The two should not be priced as if they were the same thing.

The scarce asset in AI infrastructure is the connection and the long-lead equipment behind it, not the chip, and that constraint has already reached South East Asia.

US LNG: what attribute-scarcity can do

US LNG shows the mechanism cleanly, though it is illustration rather than proof the AI-power trade plays out the same way. US liquefaction capacity reached roughly 11.4 billion cubic feet a day by the end of 2022, making the United States the world's largest LNG exporter from the first half of that year.¹⁰ Terminals are long-lead, hard-to-permit physical assets financed on long-term offtake. When spot gas prices spiked in 2021-22, physical terminal limits meant only the holders of in-service, contracted capacity could monetise the move; a buyer with capital but no terminal could not manufacture export capacity on that timeline. That is the inventory-and-lead-time rhyme: a scarce, slow-to-build physical input captures durable value across a cycle. The analogy breaks on the specifics. LNG sells into a liquid, decades-old contract market, while the AI-power constraint is being priced in real time and could clear faster than LNG capacity ever did. It points at where value should sit, not at a settled market.

How the market is repricing the constraint in real time

The capacity market is already paying for the bottleneck. PJM's 2027/28 base residual auction, announced on 17 December 2025, cleared at the regional price cap of USD 333.44 per MW-day, a record for the third auction running, at a total cost of about USD 16.4bn.⁶ The auction fell roughly 6,625 MW below PJM's 20% reserve-margin target, the first time the whole system has come up short, and the demand-forecast increase that drove it was almost entirely data centres.⁶ A clearing price at the cap tells you delivered capacity is scarce; it does not add a transformer or shorten a queue. The risk for a holder is to capitalise a cyclical clearing spread as if it were a durable, structural edge. The durable edge is the connection and the kit; the price spike is the market noticing.

The constraint travels with the buildout

The same bottleneck reappears wherever the buildout lands. Global data-centre electricity use grew about 17% in 2025, with AI-focused load up around 50%, and the IEA's base case has demand reaching about 945 TWh by 2030, with AI load set to roughly triple (the 2030 figure is an IEA projection, not an outturn).⁷ In South East Asia the constraint is already concrete. Johor's approved and under-construction data-centre pipeline reached about 5.7 GW by mid-2025, against Singapore's roughly 1.6 GW, and Malaysia plans up to 8 GW of new gas-fired power largely to serve it.⁸ Vietnam's revised Power Development Plan VIII, approved on 15 April 2025, targets a near-doubling of capacity by 2030, with electricity demand growing at around 7% a year.⁹ The question across each market is the same one we apply to SE Asia power generally: who can actually deliver capacity, and who locally holds the scarce connection. We treat that financial-close test, rather than the signing ceremony, as the real measure of a regional power project, the same lens we set out in the waste-to-energy financing note and on SE Asia industrial and logistics real estate.

Capital can buy a campus and a year of chips long before it can buy a connected megawatt.

Base, better, and worse: how the read could run

Our base case is that the equipment-and-interconnection bottleneck persists through the buildout. Transformer and turbine lead times stay stretched, gas-turbine slots fill years ahead, and the US queue median holds near 55 months, so pricing power stays structural for holders of the connection and the kit while latecomers wait. The better case for a thesis-holder is that demand and the bottleneck intensify faster than supply can answer: hyperscaler capex steps up as guided while queue reform and equipment capacity lag, SE Asia demand compounds, and the gap between an in-service connection and a fresh queue entry becomes the dominant value driver across the chain. The worse case is the falsifier firing: lead times compress back toward pre-2020 norms of 24 to 30 months, queue reform sharply cuts the median time to commercial operation, and behind-the-meter generation lets data centres bypass the grid at scale, so delivered power stops being the gating item. Or demand itself rolls over and the kit becomes stranded inventory. These figures move every quarter; the freshest readings here are within roughly six months, and most within ninety days, so anything older than about two quarters needs a refresh.

The strongest case against this read

The sharpest counter is that lead-time scarcity is the most cyclical thing in the chain dressed up as structural, and that capital plus engineering will clear it faster than we assume. It has three legs, and each is serious. Queue reform was built for exactly this: the US regulator's Order 2023, issued in July 2023, mandates first-ready, first-served cluster studies for the generation-interconnection process, and PJM's first reformed transition cycle, finalised on 20 November 2025, studied around 40 GW and issued draft agreements for about 17 GW.¹¹ Equipment supply responds to record demand, so lead times that stretched on a demand shock can compress when capacity catches up. And the grid can be bypassed: a dedicated behind-the-meter gas plant can be built in around 18 months, and midstream players are pivoting to on-site power.¹² On top of all three, the edge is demand-contingent, and a capex roll-over would strand the long-lead position.

Where the thesis holds is in how the position is held. The honest reply is not that the bear is wrong, but that each leg is observable in advance. Reform is real but slow, and it works on the generation queue that feeds the same congested grid: PJM's first reformed cycle cleared a fraction of what applied, and the median time to service is still approaching five years.⁴¹¹ Behind-the-meter is intent, not yet a delivered bypass: of roughly 90 GW of identified projects, only about 2 GW is operating today, and operators frame on-site gas as a bridge to the grid, not a replacement.¹² So the falsifier is testable rather than a matter of faith, which is why the read is held as a multi-year operator position in the scarce input, sized to evidence, not as a permanent-scarcity sector call. We hold it alongside the demand-side question we set out separately, whether headline strength is durable or rented, in consumer resilience.

What would change this view

The read breaks if delivered power stops being the gating item before the buildout completes. Concretely: transformer and gas-turbine lead times compressing back toward 24 to 30 months; interconnection-queue reform sharply cutting the median time to commercial operation from today's roughly 55 months; and on-site or behind-the-meter generation letting data centres bypass the grid at scale. It also breaks on the demand side if the AI capex cycle rolls over and demand for the long-lead kit collapses. The signposts follow directly: the next LBNL queue edition and the throughput of Order 2023 reform; the recurring transformer and turbine lead-time surveys; gas-turbine backlog and whether reservations stay sold out through the end of the decade; capacity-auction clears as a real-time gauge; large-scale behind-the-meter announcements; and SE Asia power commitments against the regional pipeline.

For diligence on any asset or name, the questions are practical. Which actually hold the scarce input, an in-service connection, a reserved turbine slot, a placed transformer order, a permitted brownfield with power, versus those still in the queue or dependent on the spot equipment market? Is the pricing power structural or a cyclical spread the cycle will hand back? How wide is the gap between an incumbent holder and a fresh queue entry in the relevant market, whether a US regional system, Johor, or Vietnam? And how exposed is the case to the falsifier set, with which indicator showing first? The decisive number, the actual clearing pace of in-service connections and equipment deliveries against the buildout, is not cleanly published by anyone. That is precisely why the position is sized to what can be proven, not to the constraint as it is sold.

Notes

1. The International Energy Agency frames the largest technology firms' capital spending as above USD 400bn in 2025, with a rise of roughly 75% expected in 2026; this is the "all large tech" cut, broader than the four-hyperscaler figure in [2]. Estimate / official framing. International Energy Agency, Energy and AI / data-centre demand analysis, 2025-2026. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
2. The four largest hyperscalers (Amazon, Alphabet, Microsoft, Meta) are guiding toward roughly USD 630-725bn of 2026 capital spending, up from about USD 380-410bn in 2025, a rise of roughly 55-77%. Company FY2026 guidance is primary; the cross-company sum is analyst aggregation (Tier-2), and overlaps with the broader IEA "all large tech" cut in [1] rather than confirming it independently. Estimate. Company FY2026 capex guidance, as of Q1 2026; aggregated by DataCenterDynamics, CNBC, Yahoo Finance, Tom's Hardware, Futurum. https://www.datacenterdynamics.com/en/opinions/tech-giants-capital-spending-surging-to-700-billion-amid-robust-ai-demand/
3. Large power transformers ran about 128 weeks (around 2.5 years) and generator step-up units about 144 weeks (around 2.8 years) in a Q2 2025 industry survey, against 24-30 months pre-2020, with high-capacity units reported up to four to five years; roughly 80% of large power transformers used in the US are imported. Reported qualitatively as a range; trade-press / survey, not a hard primary figure. POWER Magazine and pv magazine USA, citing an industry survey and PwC analysis; reporting May 2026. https://pv-magazine-usa.com/2026/05/11/u-s-transformer-market-faces-severe-supply-constraints-as-lead-times-extend-to-four-years/
4. Generators reaching commercial operation in the US in 2024 spent a median of about 55 months (around 4.5 years) in the generation-interconnection queue, up from about 22 months in 2008. Of capacity that entered queues 2000-2019, only about 19% of projects and 13% of capacity had reached commercial operation by the end of 2024, with roughly 77% of that capacity withdrawn and the balance still active or under construction. This is the generation queue, not a separate data-centre load queue; both clear the same study process and congested grid. Fact, primary/official; as of end-2024. Lawrence Berkeley National Laboratory, Queued Up: 2025 Edition (released 15 December 2025). https://emp.lbl.gov/publications/queued-2025-edition-characteristics
5. About 2,290 GW of generation and storage capacity was active in US interconnection queues at the end of 2024, down roughly 12% year-on-year as some 700 GW withdrew against about 500 GW of new requests. The widely-cited figure of about 2,600 GW is the prior 2024 Edition (as of end-2023) and is superseded here. Fact, primary/official; as of end-2024 (2025 ed.). Lawrence Berkeley National Laboratory, Queued Up: 2025 Edition. https://emp.lbl.gov/queues
6. PJM's 2027/28 base residual auction, announced 17 December 2025, cleared at the USD 333.44 per MW-day regional price cap, a record for the third consecutive auction, at a total cost of about USD 16.4bn (up from about USD 16.1bn), and fell roughly 6,625 MW below PJM's 20% reserve-margin target. The demand-forecast increase of about 5,250 MW was almost entirely data centres (about 5,100 MW). Volume procured per PJM's release is 134,479 MW on an unforced-capacity (UCAP) basis; a 145,777 MW figure in secondary reporting does not reconcile from the public releases, so the primary PJM number is used here. Fact, primary/official; auction 17 December 2025. PJM Interconnection news release (17 December 2025); secondary via Utility Dive. https://www.utilitydive.com/news/pjm-interconnection-capacity-auction-data-center/808264/
7. Global data-centre electricity use grew about 17% in 2025, with AI-focused load up around 50%; the IEA base case sees demand reaching about 945 TWh by 2030 (roughly doubling, just under 3% of global electricity), with AI-focused load set to roughly triple. The 2030 figure is an IEA projection (estimate); the 2025 growth is official. International Energy Agency, Energy and AI / data-centre demand analysis, 2025-2026. https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
8. Johor's approved and under-construction data-centre pipeline reached about 5.7 GW by mid-2025, against Singapore's roughly 1.6 GW, and Malaysia plans up to 8 GW of new gas-fired power by 2030 largely to serve data centres. Pipeline figures Tier-2 (think-tank/reporting); the 8 GW gas-power plan reported qualitatively. As of mid-2025 pipeline; 2026 reporting. Asia Society Policy Institute; regional reporting. https://asiasociety.org/policy-institute/malaysias-gamble-turning-data-centres-industrial-power
9. Vietnam's revised Power Development Plan VIII (Decision 768, approved 15 April 2025) targets installed capacity of about 183-236 GW by 2030, a near-doubling from the prior plan's 2030 level, with electricity demand growing at around 7% a year over the plan horizon. Fact / official policy target; the 2030 capacity is a planned target. Vietnam revised Power Development Plan VIII (Decision 768), via Vietnam Briefing / KPMG summaries. https://www.vietnam-briefing.com/news/vietnam-revises-pdp8-key-targets-of-the-national-power-development-plan.html/
10. US LNG liquefaction capacity reached roughly 11.4 billion cubic feet a day nominal (peak around 13.9 Bcf/d) by the end of 2022, making the United States the world's largest LNG exporter from the first half of 2022, ahead of Australia and Qatar. Terminals are financed on long-term offtake (commonly around 20-year terms), so when spot prices spiked in 2021-22 only holders of in-service capacity could monetise the move. Fact for capacity and largest-exporter status (EIA, primary/official); contract-structure detail qualitative (Columbia CGEP / CSIS). Capacity as of end-2022. US Energy Information Administration. https://www.eia.gov/todayinenergy/detail.php?id=50598
11. GE Vernova's gas-power backlog plus slot-reservation agreements grew from 83 GW to about 100 GW in Q1 2026 (around USD 4bn of Q1 gas orders), targeting at least 110 GW by year-end 2026; with about 10 GW of annual turbine capacity through 2030, the book is roughly a decade of output, and the chief executive expects reservations to be sold out through 2030 by the end of 2026 (CEO guidance/expectation, not a booked fact). On reform, the US regulator's Order 2023 (issued 28 July 2023; 2023-A on rehearing 21 March 2024) mandates first-ready, first-served cluster studies for the generation-interconnection process; in PJM's first reformed transition cycle (finalised 20 November 2025) about 40 GW of the more than 100 GW that applied were studied and about 17 GW received draft interconnection agreements. Fact for the booked backlog and the Order dates (primary); the sold-out-through-2030 expectation and reform throughput are estimate/reporting. GE Vernova Q1 2026 Form 8-K (released 22 April 2026); FERC Order 2023 explainers; PJM Inside Lines / Modo Energy. https://www.sec.gov/Archives/edgar/data/0001996810/000199681026000063/gevpressrelease1q26.htm
12. A dedicated behind-the-meter gas plant can be built in roughly 18 months against multi-year queue waits, and midstream players are pivoting to on-site power; but of about 90 GW of identified behind-the-meter projects, only around 2 GW (about 2.2%) is operating today, growing to roughly 2.8-3.2 GW by end-2026, and operators frame on-site gas as a bridge to the grid rather than a permanent replacement. Estimate / trade-press, reported as ranges; reporting late 2025-2026. Natural Gas Intelligence; Cleanview behind-the-meter report; SemiAnalysis; datacenterHawk. https://cleanview.co/reports/behind-the-meter-data-centers